CVRPLib: CVRP Benchmark Repository & Competitions

• CVRPLib Website is a specialized online platform that hosts a repository of systematically generated CVRP instances and facilitates reproducible benchmarking and live competition challenges.
• It integrates diverse benchmark families, including the new XL set, enabling empirical evaluations of both classical and machine learning-based solution methods.
• The platform offers practical tools such as map visualizations, bulk downloads, and Python generator scripts to support scalable research and direct algorithm comparisons.

CVRPLib Website is a specialized online platform serving as a repository, benchmark suite, and competition portal for the Capacitated Vehicle Routing Problem (CVRP). It provides access to systematically generated CVRP instances, solution metadata, and live benchmarking tools. The site is widely used for the experimental evaluation and direct comparison of solution methods, now including the recently introduced large-scale "XL" benchmark family. Its infrastructure supports reproducibility, community-driven best-known-solution tracking, and empirical studies of algorithmic performance, serving as a reference for both classical and machine learning-based approaches to CVRP (Queiroga et al., 16 Jan 2026).

1. Site Architecture and User Workflow

CVRPLib is structured around a central homepage (https://galgos.inf.puc-rio.br/cvrplib/), which presents a menu with direct access to major CVRP instance families: ABE­FGMPT, X, XML, XL, AGS, AV. Upon selecting "XL Instances," a user is presented with a tabular index of all 100 XL problems, showing columns for Name, Depot, Customer-positioning, Demand distribution, Capacity $Q$, average route size $r$, initial best-known-solution (BKS) value, and Method. Each entry provides links to an instance detail page with:

• A map visualization of the instance geometry.
• Direct download of the raw instance in TSPLIB-style format (coordinate list, demands, $Q$).
• Current BKS value and attribution.

Bulk resources are available, including a ZIP archive for all XL instances and the Python generator script for local reproduction or extension of the set. The website further hosts a "Best‐Known‐Solution Challenge" portal (https://galgos.inf-puc-rio.br/cvrplib/en/bks_challenge/overview) featuring:

• An overview of competition rules.
• A form for uploading routed solutions.
• Two real-time leaderboards: per‐instance (showing chronological BKS evolution) and global (ranking by "lead‐time" score).

This design enables researchers to benchmark algorithms, submit improvements, and track their comparative standings instantaneously (Queiroga et al., 16 Jan 2026).

2. XL Instance Family and Underlying Mathematical Model

The XL instance set extends previous "X" benchmarks (100–1,000 customers) to scale up to 10,000 customers, addressing the previously sparse interval of 1,000–10,000. All XL instances are Euclidean, with integer coordinates in $[0, 1000]^2$, and are generated by the same principles as Uchoa et al. (2017):

• Depot Location: Random (R), Central ($C = (500,500)$), or Eccentric ($E = (0,0)$).
• Customer Layout: Random (R), Clustered (C; $k \sim \mathrm{UD}[2,6]$ with exponential attraction), or RC (mixture).
• Demand Distribution: U (unitary), $\mathrm{UD}[1,10]$, $\mathrm{UD}[5,10]$, $\mathrm{UD}[1,100]$, $\mathrm{UD}[50,100]$, Quadrant-based (Q: $\mathrm{UD}[1,50]$ or $\mathrm{UD}[51,100]$ per quadrant), SL (mostly $\mathrm{UD}[1,10]$ with few $\mathrm{UD}[50,100]$).
• Average Route-size $r$: Ranging from Ultra short ($\mathrm{U}[3,5]$) to Ultra long ($\mathrm{U}[50,200]$).

Each instance name follows "XL-nA-kB," where $A$ gives the number of points (customers + depot) and $B$ is $K_{\min}$, the lower bound on number of routes from an exact bin-packing subproblem. Solutions with more than $K_{\min}$ routes are permitted.

Mathematically, each instance specifies a complete graph $G=(V,E)$ (with $|V|=A$), integer demand $q_j$ for each customer, and vehicle capacity $Q$. The classical CVRP is formulated as:

• Variables: $x_{ij}^k \in \{0,1\}$ indicating vehicle $k$ traverses arc $(i,j)$.
• Objective: $\min \sum_k \sum_{(i,j)\in E} d_{ij}\, x_{ij}^k$.
• Constraints:
  • $\sum_k \sum_j x_{ij}^k = 1$ for each customer $i$.
  • $\sum_j q_j\, x_{ij}^k \leq Q$ for each $k$.
  • Subtour elimination (via flow/cut-set constraints).
  • Depot entry/exit bounds: $x_{0j}^k \leq 1, x_{i0}^k \leq 1$.

All data are integral, yielding large-scale mixed-integer programs for empirical evaluation (Queiroga et al., 16 Jan 2026).

3. Computational Study and BKS Challenge Mechanism

Prior to the public launch of the XL BKS Challenge (January 12, 2026), initial BKS values were established by executing seven state‐of‐the‐art algorithms on each instance, with 60 independent two-hour runs per instance using a single AMD EPYC 9654 core. This resulted in 120 CPU hours per instance. The algorithms evaluated included:

• AILS-II (Java)
• FILO, FILO2 (C++)
• KGLS^XXL (Java)
• SISRs (C++)
• HGS-CVRP (C++)
• LKH-3 (C)
• Google OR-Tools (C++)

AILS-II found 93 of 100 initial BKSs, FILO2 accounted for 6, FILO for 1. Observed mean percentage gaps to BKS were 0.07% (AILS-II), 0.21% (FILO2), 0.25% (FILO), 1.00% (KGLS^XXL), and over 1% for the remaining methods.

BKS Challenge submissions are automatically checked for feasibility and objective value. Each strictly better feasible solution updates the instance's BKS and attributes a "lead-time" reward (days unbeaten) to the submitting team, with a +5-day bonus for the final contest BKS. Leaderboards provide real-time feedback per-instance and globally by aggregate lead-time (Queiroga et al., 16 Jan 2026).

4. Integration with Existing CVRPLib Benchmarks

CVRPLib also hosts the classic X set and the ML-oriented XML100 set, each with BKS tracking and visualization. With the inclusion of XL, the website comprehensively covers the benchmark spectrum:

• Scale Coverage: XL spans 1,000–10,000, bridging the former gap between X (100–1,000), AGS (3,000–30,000), and AV (20,000–1,000,000; less instance diversity).
• Algorithm Performance Trends: HGS-CVRP excels on X, while AILS-II is superior on XL. On XML100 (10,000 × 100-customer), all classic methods reach optimality in under one minute, critical for error analysis in ML contexts.
• Comparative Navigation: Users can view tables for X, XML, and XL, comparing best and mean gaps as instance size increases.

This suggests the platform enables integrated, repeatable algorithm benchmarking over multiple regimes of instance size and structure (Queiroga et al., 16 Jan 2026).

5. Technical Resources and URLs

The primary resources and entry points are as follows:

Purpose	CVRPLib Resource
Main site / navigation	https://galgos.inf.puc-rio.br/cvrplib/
XL instance download and description	https://galgos.inf.puc-rio.br/cvrplib/en/instances/xl
Python instance generator script	Linked as “XL generator script” from XL instances page
BKS Challenge portal/rules/timeline	https://galgos.inf-puc-rio.br/cvrplib/en/bks_challenge/overview
Solution submission form	Accessed from "Submit" button on Challenge overview page
Per-instance and global leaderboards	Tabs within Challenge overview

These tools and interfaces support direct download, local reproduction, benchmarking, and participation in ongoing community driven optimization challenges.

6. Significance and Impact

CVRPLib’s extension with the XL benchmark and associated infrastructure establishes a unified, scalable repository for the systematic advancement of CVRP computational research. It enables both incremental improvement of best-known results and transparent assessment of new algorithms at scale. By integrating reproducibility tools, scalable instance generation, and live leaderboards, it sets a consistent experimental standard for the vehicle routing community and supports a wide spectrum of methodological approaches, including metaheuristics and learning-based solvers (Queiroga et al., 16 Jan 2026).